C++ GPT Runtime

TensorRT-LLM includes a C++ component to execute TensorRT engines built with the Python API as described in the TensorRT-LLM Architecture section. That component is called the C++ runtime.

The API of the C++ runtime is composed of the classes declared in cpp/include/tensorrt_llm/runtime and implemented in cpp/tensorrt_llm/runtime.

Even if the different components described in that document mention GPT in their name, they are not restricted to this specific model. Those classes can be used to implement auto-regressive models like BLOOM, GPT-J, GPT-NeoX or LLaMA, for example.

Complete support of encoder-decoder models, like T5, will be added to TensorRT-LLM in a future release. An experimental version, only in Python for now, can be found in the examples/enc_dec folder.

Overview

Runtime models are described by an instance of the ModelConfig class and a pointer to the TensorRT engine that must be executed to perform the inference. The environment is configured through the WorldConfig (that name comes from MPI and its “famous” MPI_COMM_WORLD default communicator). The SamplingConfig class encapsulates parameters that control the generation of new tokens.

Model Configuration

The model configuration is an instance of the ModelConfig class. That class encapsulates the following parameters (they are declared as private member variables and exposed through getters and setters):

vocabSize, the size of the vocabulary,
numLayers, the number of layers in the model,
numHeads, the number of heads in the attention block,
numKvHeads, the number of heads for K and V in the attention component. When the number of K/V heads is the same as the number of (Q) heads, the model uses multi-head attention. When the number of K/V heads is 1, it uses multi-query attention. Otherwise, it uses group-query attention. Refer to Multi-Head, Multi-Query, and Group-Query Attention for more information,
hiddenSize, the size of the hidden dimension,
dataType, the datatype that was used to build the TensorRT engine and that must be used to run the model during inference,
useGptAttentionPlugin, indicates if the Multi-Head, Multi-Query, and Group-Query Attention operator was compiled using the GPT Attention plugin,
inputPacked, indicates that the input must be packed (or padded when set to false). For performance reasons, it is recommended to always use packed, even if its default is set to false (will be changed in a future release). Refer to Multi-Head, Multi-Query, and Group-Query Attention for more information,
pagedKvCache, indicates if the K/V cache uses paging. Refer to Multi-Head, Multi-Query, and Group-Query Attention for more information,
tokensPerBlock, is the number of tokens in each block of the K/V cache. It’s relevant when the paged K/V cache is enabled. By default, the value is 64. Refer to Multi-Head, Multi-Query, and Group-Query Attention for more information,
quantMode, controls the quantization method. Refer to Numerical Precision for more information.
maxBatchSize, indicates the maximum batch size that the TensorRT engine was built for,
maxInputLen, the maximum size of the input sequences,
maxSequenceLen, the maximum total size (input+output) of the sequences.

World Configuration

Familiarity with MPI, is not required to utilize the TensorRT-LMM C++ runtime. There are two main things you need to know:

The C++ Runtime in TensorRT-LLM uses processes to execute TensorRT engines on the different GPUs. Those GPUs can be located on a single node as well as on different nodes in a cluster. Each process is called a rank in MPI.
The ranks are grouped in communication groups. The TensorRT-LLM C++ Runtime calls that group the world.

The world configuration is an instance of the WorldConfig class, which encapsulates the following parameters:

tensorParallelism, the number of ranks that collaborate together to implement Tensor Parallelism (TP). With TP, each GPU performs computations for all the layers of the model. Some of those computations are distributed across the GPU. TP is more balanced than Pipeline Parallelism (PP), in most cases, but requires higher bandwidth between the GPUs. It is the recommended setting in the presence of NVLINK between GPUs,
pipelineParallelism, the number of ranks that collaborate together to implement Pipeline Parallelism (PP). With PP, each GPU works on a subset of consecutive layers. Communications between the GPUs happen only at the boundaries of the subsets of layers. It is harder to guarantee the full utilization of the GPUs with PP but it requires less memory bandwidth. It is the recommended setting in the absence of NVLINK between GPUs,
rank, the unique identifier of the rank,
gpusPerNode, indicates the number of GPUs on each node. Having that information allows the C++ runtime to optimize communications between GPUs in a node (like taking advantage of the NVLINK interconnect between GPUs of an A100 DGX node).

Sampling Parameters

The SamplingConfig class encapsulates parameters that control the generation of new tokens. A comparison of selecting decoding method is listed as the table below (X means it is not supported yet). Except for the beamWidth parameter, all the fields are optional and the runtime will use a default value if no values are provided by the user. For vector fields, the TensorRT-LLM runtime supports one value per sequence (that is, the vector contains batchSize values). If all the sequences use the same value for a given parameter, the vector can be limited to a single element (that is, size() == 1).

Method name in HF	Condition in HF	Method name in TRT-LLM	Condition in TRT-LLM
assisted decoding	`assistant_model` or `prompt_lookup_num_tokens!=None`	X
beam-search decoding	`num_beams>1` and `do_sample=False`	beam search	`beamWidth > 1`
beam-search multinomial sampling	`num_beams>1` and `do_sample=True`	X
constrained beam-search decoding	`constraints!=None` or `force_words_ids!=None`	X
contrastive search	`penalty_alpha>0` and `top_k>1`	X
diverse beam-search decoding	`num_beams>1` and `num_beam_groups>1`	X
greedy decoding	`num_beams=1` and `do_sample=False`	sampling	`beamWidth == 1` and `topK=0` and `topP=0.0f`
multinomial sampling	`num_beams=1` and `do_sample=True`	sampling	`beamWidth == 1` and (`topK>0` or `topP>0.0f`)

General

Name in TRT-LLM	Description	Data type	Range of value	Default value	Name in HF
`temperature`	modulation of logits in sampling workflow	List[Float]	[0.0f, $+\infty$)	`1.0f` (no modulation)	`temperature`
`minLength`	lower-bound on the number of tokens generated	List[Int]	[0, $+\infty$)	`0` (no effect (the first generated token can be EOS)	`min_length`
`repetitionPenalty`	penalize repetitive tokens multiplicative, irrespective of appearances count	List[Float]	[0.0f, $+\infty$) `< 1.0f` encourages repetition `> 1.0f` discourages it	`1.0f` (no effect)	`repetition_penalty`
`presencePenalty`	penalize existed tokens additive, irrespective of appearances count	List[Float]	($-\infty$, $+\infty$) `< 0.0f` encourages repetition `> 0.0f` discourages it	`0.0f` (no effect)	no
`frequencyPenalty`	penalize existed tokens additive, dependent on appearances count	List[Float]	($-\infty$, $+\infty$) `< 0.0f` encourages repetition `> 0.0f` discourages it	`0.0f` (no effect)	no
`noRepeatNgramSize`		List[Int]	[0, $+\infty$) `> 0` all ngrams of that size can only occur once	`0` (no effect)	`no_repeat_ngram_size`

The tokens of input prompt are included during adopting repetitionPenalty, presencePenalty, and frequencyPenalty onto logits.
The parameters repetitionPenalty, presencePenalty, and frequencyPenalty are not mutually exclusive.

Sampling

Name in TRT-LLM	Description	Data type	Range of value	Default value	Name in HF
`randomSeed`	random seed for random number generator	Int64	[0, 2^64-1]	`0`	no
`topK`	the number of logits to sample from	List[Int]	[0, 1024]	`0`	`top_k`
`topP`	the top-P probability to sample from	List[Float]	[0.0f, 1.0f]	`0.0f`	`top_p`
`topPDecay`	the decay in the `topP` algorithm	List[Float]	(0.0f, 1.0f]	`1.0f`	no
`topPMin`	the decay in the `topP` algorithm	List[Float]	(0.0f, 1.0f]	`1.0e-6,f`	no
`topPResetIds`	the decay in the `topP` algorithm	List[Int]	[-1, $+\infty$)	`-1` (no effect)	no

If setting topK = 0 and topP = 0.0f, greedy search is performed.
If setting topK > 0 and topP = 0.0f, topK tokens of highest probabilities will become the candidates of sampling (named TopK sampling in TRT-LLM).
If setting topK = 0 and topP > 0.0f, tokens will be sorted with probability descendly, then the tokens with highest probabilities which the accumulated probability larger than topP will become the candidates of sampling (named TopP sampling in TRT-LLM).
If setting topK > 0 and topP > 0.0f, topK tokens of highest probabilities will be selected, then those selected tokens will be sorted with probability descendly and their probability will be normalized, then the tokens with highest normalized probabilities which the accumulated probability larger than topP will become the candidates of sampling (named TopKTopP sampling in TRT-LLM)
If different topK values are provided for the different sequences in the batch, the performance of the implementation will depend on the largest value. For efficiency reasons, we recommend to batch requests with similar topK values together.
topPDecay, topPMin and topPResetIds are explained in Factuality Enhanced Language Models for Open-Ended Text Generation. topPDecay is the decay, topPMin is the lower-bound and topPResetIds indicates where to reset the decay.

Beam-search

Name in TRT-LLM	Description	Data type	Range of value	Default value	Name in HF
`beamWidth`	width for beam-search algorithm	Int	[0, 64]	`0` (disable beam search)	`beam_width`
`beamSearchDiversityRate`	diversity of generated tokens	List[Float]	[0, $+\infty$)	`0.0f`	`diversity_penalty`
`lengthPenalty`	penalize longer sequences	List[Float]	[0, $+\infty$)	`0.0f`	`length_penalty`
`earlyStopping`	see description below	List[Int]	($-\infty$, $+\infty$)	`0`	`early_stopping`

Beam-search algorithm: beam search.
Parameter diversity_penalty in HF is only used for diverse beam-search decoding (or named Group-Beam-Search), which is not supported by TRT-LLM yet.
If setting earlyStopping = 1, decoding will stop once beamWidth finished sentences are generated.
If setting earlyStopping = 0, decoding will keep going until no better sentences (with better score) can be generated.
If setting earlyStopping to other values, decoding will stop only depending on lengthlengthPenalty.
The beamWidth parameter is a scalar value. It means that in this release of TensorRT-LLM, it is not possible to specify a different width for each input sequence. This limitation is likely to be removed in a future release.

The Session

The runtime session is deprecated in favor of the Executor API. It will be removed in a future release of TensorRT-LLM.

An example of how to use the GptSession to run a GPT-like auto-regressive model can be found in cpp/tests/runtime/gptSessionTest.cpp.

Internal Components

The GptSession class encapsulates two main components. The TllmRuntime is in charge of the execution of the TensorRT engine. The GptDecoder does the generation of the tokens from the logits. The TllmRuntime class is an internal component and you are not expected to use that class directly. The GptDecoder can be used directly to implement custom generation loop and for use cases that cannot be satisfied by the implementation in GptSession.

In-flight Batching Support

In-flight batching is supported using separate decoders per request. The biggest difference compared to using a single decoder is in how the token generation from logits is managed. A batch is split into batchSize individual requests and kernels are issued using separated CUDA streams. This behavior may be revisited in a future release to maintain the structure of the batch and improve efficiency.

Know Issues and Future Changes

In the current release of TensorRT-LLM, the C++ and Python runtimes are two separate software components and the C++ runtime is being more actively developed (with features like in-flight batching). An objective, for a future release, could be to rebuild the Python runtime on top of the C++ one.